RGB-Based 3D Reconstruction: An Investigation of SfM and 3D Gaussian Splatting Performances of SIFT, Superpoint, and ALIKED Methods

Abstract

The three-dimensional (3D) reconstruction of indoor scenes from two-dimensional (2D) RGB images is a fundamental research topic in the field of computer vision. In this study, the effects of classical (SIFT) and learned (SuperPoint, ALIKED) local feature extraction methods, along with learned matching (LightGlue) algorithms, on Structure-from-Motion (SfM) and the final 3D Gaussian Splatting (3DGS) performance in the indoor 3D reconstruction process are comparatively investigated. The findings demonstrate that while learned algorithms provide more consistent matches and stable camera poses, the denser sparse point cloud generated by the SIFT-based classical approach offers higher visual quality (PSNR, SSIM, LPIPS) in the 3DGS stage.